Automatic segmentation of surface EMG images: Improving the estimation of neuromuscular activity

Surface electromyograms (EMGs) recorded with a couple of electrodes are meant to comprise representative information of the whole muscle activation. Nonetheless, regional variations in neuromuscular activity seem to occur in numerous conditions, from standing to passive muscle stretching. In this study, we show how local activation of skeletal muscles can be automatically tracked from EMGs acquired with a bi-dimensional grid of surface electrodes (a grid of 8 rows and 15 columns was used). Grayscale images were created from simulated and experimental EMGs, filtered and segmented into clusters of activity with the watershed algorithm. The number of electrodes on each cluster and the mean level of neuromuscular activity were used to assess the accuracy of the segmentation of simulated signals. Regardless of the noise level, thickness of fat tissue and acquisition configuration (monopolar or single differential), the segmentation accuracy was above 60%. Accuracy values peaked close to 95% when pixels with intensity below approximately 70% of maximal EMG amplitude in each segmented cluster were excluded. When simulating opposite variations in the activity of two adjacent muscles, watershed segmentation produced clusters of activity consistently centered on each simulated portion of active muscle and with mean amplitude close to the simulated value. Finally, the segmentation algorithm was used to track spatial variations in the activity, within and between medial and lateral gastrocnemius muscles, during isometric plantar flexion contraction and in quiet standing position. In both cases, the regionalization of neuromuscular activity occurred and was consistently identified with the segmentation method